AI Agents Learn to Find Vulnerabilities in Smart Contracts

Research from Anthropic, OpenAI, and other teams indicates that AI agents are capable of independently identifying and reproducing vulnerabilities in smart contracts. In laboratory experiments, models successfully simulated attacks worth hundreds of millions of dollars. However, these results do not imply that AI has already caused comparable losses in active DeFi protocols.
Anthropic discovered new vulnerabilities using AI
In December 2025, Anthropic presented the results of the SCONE-bench study, in which AI models were tested for their ability to reproduce attacks on smart contracts. The sample included 405 contracts whose vulnerabilities were exploited by attackers from 2020 to 2025.
AI agents successfully reproduced attacks on 207 contracts. The total value of conditionally extracted funds amounted to $550.1 million. Additional testing of vulnerabilities disclosed after the control dates of model training showed that Claude Opus 4.5, Claude Sonnet 4.5, and GPT-5 are capable of reproducing exploits worth $4.6 million.
Researchers also tested 2,849 recently deployed smart contracts that had no known vulnerabilities. As a result, AI agents discovered two previously unknown security issues. In simulation, their exploitation allowed for the conditional extraction of assets worth $3,694, while API costs in one of the GPT-5 experiments amounted to $3,476.
Anthropic emphasized that all tests were conducted in isolated environments. Real blockchain networks and user funds were not affected.
Other studies confirmed the capabilities of AI agents
In February 2026, OpenAI and Paradigm introduced EVMbench – a system for testing AI agents in detecting, fixing, and exploiting smart contract vulnerabilities.
In exploitation tasks, GPT-5.3-Codex achieved a result of 71%, while GPT-5 successfully completed 33.3% of the tests. The attacks were reproduced in a local blockchain environment, not in active networks.
In October, researchers published CyberChainBench, covering 541 historical cases of vulnerability exploitation in nine blockchains compatible with Ethereum. The best AI agent configuration achieved 37.5% in vulnerability detection, 43.7% in attack reproduction, and 23.4% in problem fixing.
In a separate series of tests, the agent reproduced attacks with an estimated result of $57.4 million on a sample of 200 cases. However, these funds were also obtained exclusively in simulation.
The results of such studies significantly depend on testing conditions. For example, in an a16z experiment, the initial success rate of the agent was 50%, but after restricting access to information about already committed attacks, it dropped to 10%. Additional specialized instructions increased the rate to 70%.
Real DeFi losses do not prove AI involvement in attacks
According to Hacken, in the third quarter of 2026, the cryptocurrency industry lost about $1.18 billion due to 107 incidents. Vulnerabilities in smart contracts and blockchain protocol software accounted for 48.2% of the total damage, or approximately $569.5 million.
A separate Chainalysis study identified thefts of $36.7 million from protocols whose exploited contracts did not have published and verified source code in blockchain explorers. This does not mean that all such contracts necessarily lacked independent audits.
Experts note that the development of reverse engineering tools and AI capabilities may simplify the search for vulnerabilities even in contracts whose source code is not published.
However, the provided data does not establish a direct link between real cryptocurrency thefts and the actions of autonomous AI agents.
How AI can impact blockchain security
The development of AI creates additional risks for DeFi, as automated systems can check a large number of smart contracts and sequentially search for errors in their logic. At the same time, such tools can be used by developers to enhance the security of their own projects.
AI agents can assist in testing contracts before deployment, preparing regression tests, and identifying unusual behavior in blockchain transactions. However, discovered issues should be reviewed by specialists, and critical changes should be confirmed by authorized team members.
Independent audits, restricting access to critical functions, continuous monitoring, and pre-prepared response plans remain important elements of DeFi protocol protection.
Thus, research confirms the growing capabilities of AI in finding and exploiting vulnerabilities. However, laboratory results cannot yet be directly applied to the scale of real attacks or used as evidence of AI involvement in major cryptocurrency thefts.
This material is prepared solely for informational purposes and does not constitute financial advice or a recommendation.




